Open Collections

Machine learning estimation of snow water equivalent over British Columbia

University of British Columbia

In regions with hydrologic regimes dominated by spring snowmelt, reliable water resources forecasting and planning require accurate estimates of snow water equivalent (SWE). While various gridded data products can provide such estimates, they are especially challenging under conditions of mountainous terrain, heavy forest cover and large snow accumulations. These conditions aptly describe the province of British Columbia (BC), Canada. SWE values from reanalyses, land data assimilation systems and observation-based data sets were compared with manual snow surveys over the five physiographic regions of BC. Yearly time series were generated for each survey month (January through June), and correlation, bias and mean absolute error (MAE) were found for each product and physiographic region. Better SWE estimates were seen in regions of lower snow accumulation and land relief, and a product performance ranking found three products to be the best: ERA-Interim/Land, GLDAS-2 and MERRA. Using the snow surveys as target data, combinations of the gridded SWE products and primary spatiotemporal covariates (survey date, year, latitude, longitude, elevation and grid cell elevation differences) were incorporated into multiple linear regression (MLR) and artificial neural network (ANN) models. The ANN using the best three products was found to have the lowest overall MAE compared to other products and models. This base ANN was then enhanced to include terrain covariates (slope, aspect and Terrain Ruggedness Index) and a simple one-layer energy balance snow model driven by bias-corrected ANUSPLIN temperatures and precipitations. Lagged predictor values compensated for early snow model melt off. These enhancements further improved mean station MAE by 10% and correlations by 0.04. The ANN models were also compared with the Variable Infiltration Capacity model calibrated for four BC watersheds. While the base ANN produced statistically significant MAE reductions in one region and correlation increases in two regions, the enhanced ANN achieved larger improvements that were significant in three and all five regions, respectively. These results demonstrate the promise of machine learning in fusing complementary information from sophisticated lower-resolution gridded products and a simple but higher-resolution snow model, improving SWE estimation in challenging contexts.

Text

2019-08-31

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/64360

Geological Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/